Existing machine learning–based predictive control systems for chamber pressure face limitations such as improper reference setting, weak dynamic adaptability, and suboptimal parameter tuning. This study proposes a predictive control framework that integrates an adaptively updated LightGBM (AU-LightGBM) with an improved grey wolf optimizer (IGWO). First, theoretical analysis combined with LightGBM is used to estimate the ultimate face support pressure, which serves as the target chamber pressure. Second, an adaptive updating strategy is embedded into LightGBM to maintain stable prediction accuracy during continuous tunneling. Third, a KD-tree is employed to retrieve high-quality historical parameter sets, which are then used to initialize the grey wolf optimizer and enhance iterative optimization. The case study results demonstrate that the proposed framework exhibits strong effectiveness and adaptability across the conventional loess stratum and water-rich sand layer of the Shaolingyuan Tunnel, as well as the complex stratigraphic conditions of the Tianjin metro project. The average coefficient of determination (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>), root mean square error (RMSE), and mean absolute error (MAE) for chamber pressure prediction are 0.943, 0.082, and 0.055, respectively. After optimization, the average <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>, RMSE, and MAE between the predicted and reference chamber pressures are 0.753, 0.143, and 0.116, respectively. Under cohesive and non-cohesive soil conditions, the MAE between the results obtained by the hybrid LightGBM model and the experimental values are 11.987 and 7.958, respectively, which are lower than those of most theoretical models. Compared with six baseline learners, AU-LightGBM achieves the highest prediction accuracy, with <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(R^2,\)</EquationSource> </InlineEquation> RMSE, and MAE of 0.984, 0.014, and 0.010, respectively, between the predicted and actual values on the test set. By leveraging clustered historical samples, IGWO strengthens the initial population and yields more robust optimized parameters than the canonical GWO. The <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>, RMSE, and MAE between the chamber pressure optimized by IGWO and the reference values are 0.944, 0.023, and 0.015, respectively, all outperforming GWO, Genetic Algorithm, and Particle Swarm Optimization, thereby validating the superiority of the proposed method.

Predictive control of earth pressure balance in shield tunneling using a hybrid learning approach

期刊论文
出版模式：

开放获取
发表日期：

2026年4月22日

查看全文

Qiyue Wang,
Yong Zhang,
Daojiang Wei,
Jiugang Meng

摘要

Existing machine learning–based predictive control systems for chamber pressure face limitations such as improper reference setting, weak dynamic adaptability, and suboptimal parameter tuning. This study proposes a predictive control framework that integrates an adaptively updated LightGBM (AU-LightGBM) with an improved grey wolf optimizer (IGWO). First, theoretical analysis combined with LightGBM is used to estimate the ultimate face support pressure, which serves as the target chamber pressure. Second, an adaptive updating strategy is embedded into LightGBM to maintain stable prediction accuracy during continuous tunneling. Third, a KD-tree is employed to retrieve high-quality historical parameter sets, which are then used to initialize the grey wolf optimizer and enhance iterative optimization. The case study results demonstrate that the proposed framework exhibits strong effectiveness and adaptability across the conventional loess stratum and water-rich sand layer of the Shaolingyuan Tunnel, as well as the complex stratigraphic conditions of the Tianjin metro project. The average coefficient of determination ( \(R^2\) ), root mean square error (RMSE), and mean absolute error (MAE) for chamber pressure prediction are 0.943, 0.082, and 0.055, respectively. After optimization, the average \(R^2\) , RMSE, and MAE between the predicted and reference chamber pressures are 0.753, 0.143, and 0.116, respectively. Under cohesive and non-cohesive soil conditions, the MAE between the results obtained by the hybrid LightGBM model and the experimental values are 11.987 and 7.958, respectively, which are lower than those of most theoretical models. Compared with six baseline learners, AU-LightGBM achieves the highest prediction accuracy, with \(R^2,\) RMSE, and MAE of 0.984, 0.014, and 0.010, respectively, between the predicted and actual values on the test set. By leveraging clustered historical samples, IGWO strengthens the initial population and yields more robust optimized parameters than the canonical GWO. The \(R^2\) , RMSE, and MAE between the chamber pressure optimized by IGWO and the reference values are 0.944, 0.023, and 0.015, respectively, all outperforming GWO, Genetic Algorithm, and Particle Swarm Optimization, thereby validating the superiority of the proposed method.